Writing
Notes from the JVM. And the cloud it runs on.
Retrying Kafka messages without blocking the partitionA scheduler-backed retry topic keeps transient failures from stalling everything behind them.Exactly-once is a contract, not a checkboxKafka's transactional producer gets you atomic writes. Everything either side of the broker is still your problem.What actually fixed our Lambda cold startsProvisioned concurrency is the expensive answer. Three cheaper changes moved p99 further.Teaching a RAG assistant to say "I don't know"Retrieval quality gets the attention. Refusal behaviour is what made ours trustworthy.Tests that survive refactoringIf a green suite goes red when you move a method, the tests were describing the code, not the behaviour.Which MCP servers were actually worth buildingFive integrations, six months in. Two changed how I work; the rest were solutions looking for problems.